A looped design lets Inkling 2 Mini reason with fewer tokens

Thinking Machines says Inkling 2 Mini reuses some transformer layers to do more internal processing without generating extra text. The open model handles text and images, and the company reports strong results on agentic coding tests alongside local use on certain computers.

WTF Index NEUTRAL
◄ Terminator 1 Idiocracy 1 ►

This is a routine open-model release with reported reasoning and coding gains, without a clear lean toward control or human skill erosion.

A looped design lets Inkling 2 Mini reason with fewer tokens

Inkling 2 Mini uses repeated passes through part of its network to carry out more internal processing while producing fewer tokens. Thinking Machines describes the release as a looped transformer: rather than relying only on a longer stream of generated text, the model runs selected layers multiple times for each token.

How the looped transformer works

The model has 24 layers, and 12 of them run three times in a row. According to Thinking Machines, this structure gives Inkling 2 Mini more room to reason internally without generating more text, helping it reach an answer with fewer tokens.

That design is the defining feature of the release. Repeating layers changes how computation is used inside the model, while the visible output can remain shorter. The source does not provide a separate token comparison, so the reported benefit is best understood as the company’s account of the architecture.

A large model with a smaller active share

Inkling 2 Mini is an open mixture-of-experts model with 26.4 billion parameters. The company says 5.4 billion are active per token because of the loops. It can process both text and images and has a 1-million-token context window.

The model is licensed under Apache 2.0. Together, the open license and the ability to run it locally may make it relevant to developers and researchers who want to explore its architecture or use it outside a hosted service.

Reported coding and intelligence results

On agentic coding tests, Inkling 2 Mini leads open models below 100 billion parameters, according to the article. It scored 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2.

Artificial Analysis gives the model 20 points on its Intelligence Index. That ties it with Qwen3.8 27B, while the article says Inkling 2 Mini does so at roughly six times lower cost and four times the speed. These comparisons describe the reported index and cost and speed figures; they do not establish that the models perform identically in every use.

Where to run and try it

Inkling 2 Mini runs locally on Macs with at least 32 GB of memory and on Nvidia's DGX Spark. The weights are available through Hugging Face, where a research version with variable loops can also be downloaded. The model can also be tried in the Tinker Playground.

For users interested in efficient reasoning, the release offers a specific idea to examine: repeat some layers internally, while keeping the model open and available for local execution on supported hardware. Its published coding scores and benchmark comparison add context, while practical results will depend on the task and setup.